Heterogeneous Compute Playground

A lightweight playground for exploring performance and energy behavior across two worlds: CUDA linear algebra kernels in modern C++, and Verilog waveform analysis through Python.

Overview

This repository combines:

• CUDA experiments for core linear-algebra operations.
• Verilog simulation workflows using wave.vcd.
• Python visualization to analyze switching activity, state breakdowns, and energy-like proxies.

The goal is to build intuition for high-performance compute from both the GPU and hardware-design sides.

Features

CUDA / C++

• Modern C++ CUDA implementation.
• Vector addition and other linear-algebra building blocks.
• Designed to be lightweight, fast, and easy to extend.

Verilog / VCD analysis

• Reads simulation traces from wave.vcd.
• Extracts signal activity over time.
• Computes switching activity, utilization, and state fractions.
• Generates execution plots, breakdown plots, and heatmaps.

Python visualization

• Signal timeline plots.
• State breakdown analysis.
• Memory-vs-compute heatmaps.
• Simple energy proxy metrics based on switching activity and signal states.

Why this repo?

This project is not just about raw speed. It is about understanding how compute behaves:

• on a GPU,
• inside a digital design,
• and across time in simulation traces.

That makes it useful for performance engineering, hardware-aware ML, and energy-efficient compute exploration.

Project structure

.
├── basic/                # CUDA / C++ source code
├── verilogFPGA/            # Verilog modules
├── wave.vcd            # Simulation waveform
├── cudaenergy.py       # VCD parsing and plotting
└── output/             # Generated plots and results (Ongoing)

Python analysis

The cudaenergy.py script:

• loads wave.vcd,
• parses signals with vcdvcd,
• computes switching activity,
• estimates an energy proxy,
• and saves plots for execution and signal breakdowns.

Example usage:

/bin/python3 cudaenegy.py wave.vcd --op "TOP.op[1:0]" --y "TOP.y[7:0]"

CUDA build

mkdir build
cd build
cmake ..
make
./your_cuda_binary

Outputs

The repository can generate:

• waveform execution plots,
• state breakdown figures,
• memory/compute heatmaps,
• CUDA benchmark results.

Roadmap

• Add more CUDA linear algebra kernels.
• Extend Verilog modules and waveform cases.
• Compare signal activity across different designs.
• Add performance and energy benchmarking tables.